Data driven patient-specialized neural networks for blood glucose prediction

Diabetes is an autoimmune disease characterized by glucose levels dysfunctions. It involves continuous monitoring combined with insulin treatment. Nowadays, continuous glucose monitoring systems (CGMs) have led to a greater availability of data. These can be effectively used by machine learning techniques to infer future values of the glycaemic concentration, allowing the early prevention of dangerous states and a better optimisation of the diabetic treatment. In this work, we investigate a patient-specialized prediction model. Thus, we designed a specialized solution based on Long Short-Term Memory (LSTM) neural network. Our solution was experimentally compared with two literature approaches, respectively based on Feed-Forward (FNN) and Recurrent (RNN) neural networks. The experimental results have highlighted that our LSTM solution obtained good performance both for short- and long-term glucose level inference (60 min.), overcoming the other methods both in terms of correlation between measured and predicted glucose signal and in terms of clinical outcome.